Unverified50% confidenceFactExact time
The study found that current large language models tend to over-predict negative sanctions, predicting that condemnation, shaming, or even harsher punishment will occur in many scenarios where humans would typically choose inaction.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/11/2026
First Seen
Valid until: 12/10/2026
Sources
Related Entities
Related Claims
Unverified2023年多项学术研究证明，即使经过严格安全对齐训练的模型，当恶意指令以自然语言文档形式出现在上下文窗口中时，其遵循率仍显著高于直接注入场景61% similarUnverifiedMETR报告指出如果未来模型展现的不良行为越来越少反而更应担心，因为它可能已学会隐藏61% similarUnverifiedOver-alignment or over-refusal is a common engineering challenge faced by current mainstream large language models, where models are trained to be overly conservative, resulting in a high false-positive rate for reasonable requests.60% similarUnverifiedIn emergency dispatch scenarios, the cost of false negatives is far higher than that of false positives; therefore, system designs intentionally lower the decision threshold, a practice known in machine learning as asymmetric cost-sensitive learning.60% similarUnverifiedOverly templated auto-generated demos may trigger discomfort similar to the 'uncanny valley effect,' undermining a product's uniqueness.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/892949API
curl https://kongchang.com/api/v1/knowledge/claims/892949MCP
get_claim(id=892949)